src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <fcntl.h>
   4 #include <sys/statvfs.h>
   5 #include <sys/vfs.h>
   6 #include <unistd.h>
   7
   8 #include "acl-util.h"
   9 #include "dirent-util.h"
  10 #include "fd-util.h"
  11 #include "fileio.h"
  12 #include "fs-util.h"
  13 #include "missing_magic.h"
  14 #include "nspawn-def.h"
  15 #include "nspawn-patch-uid.h"
  16 #include "stat-util.h"
  17 #include "stdio-util.h"
  18 #include "string-util.h"
  19 #include "strv.h"
  20 #include "user-util.h"
  21
  22 #if HAVE_ACL
  23
  24 static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
  25         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  26         acl_t acl;
  27
  28         assert(fd >= 0);
  29         assert(ret);
  30
  31         if (name) {
  32                 _cleanup_close_ int child_fd = -1;
  33
  34                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  35                 if (child_fd < 0)
  36                         return -errno;
  37
  38                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  39                 acl = acl_get_file(procfs_path, type);
  40         } else if (type == ACL_TYPE_ACCESS)
  41                 acl = acl_get_fd(fd);
  42         else {
  43                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  44                 acl = acl_get_file(procfs_path, type);
  45         }
  46         if (!acl)
  47                 return -errno;
  48
  49         *ret = acl;
  50         return 0;
  51 }
  52
  53 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  54         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  55         int r;
  56
  57         assert(fd >= 0);
  58         assert(acl);
  59
  60         if (name) {
  61                 _cleanup_close_ int child_fd = -1;
  62
  63                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  64                 if (child_fd < 0)
  65                         return -errno;
  66
  67                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  68                 r = acl_set_file(procfs_path, type, acl);
  69         } else if (type == ACL_TYPE_ACCESS)
  70                 r = acl_set_fd(fd, acl);
  71         else {
  72                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  73                 r = acl_set_file(procfs_path, type, acl);
  74         }
  75         if (r < 0)
  76                 return -errno;
  77
  78         return 0;
  79 }
  80
  81 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  82         _cleanup_(acl_freep) acl_t copy = NULL;
  83         acl_entry_t i;
  84         int r;
  85
  86         assert(acl);
  87         assert(ret);
  88
  89         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  90         if (r < 0)
  91                 return -errno;
  92         while (r > 0) {
  93                 uid_t *old_uid, new_uid;
  94                 bool modify = false;
  95                 acl_tag_t tag;
  96
  97                 if (acl_get_tag_type(i, &tag) < 0)
  98                         return -errno;
  99
 100                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
 101
 102                         /* We don't distinguish here between uid_t and gid_t, let's make sure the compiler checks that
 103                          * this is actually OK */
 104                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
 105
 106                         old_uid = acl_get_qualifier(i);
 107                         if (!old_uid)
 108                                 return -errno;
 109
 110                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 111                         if (!uid_is_valid(new_uid))
 112                                 return -EINVAL;
 113
 114                         modify = new_uid != *old_uid;
 115                         if (modify && !copy) {
 116                                 int n;
 117
 118                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 119                                  * beginning, so that we copy all entries, starting from the first, this time. */
 120
 121                                 n = acl_entries(acl);
 122                                 if (n < 0)
 123                                         return -errno;
 124
 125                                 copy = acl_init(n);
 126                                 if (!copy)
 127                                         return -errno;
 128
 129                                 /* Seek back to the beginning */
 130                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 131                                 if (r < 0)
 132                                         return -errno;
 133                                 continue;
 134                         }
 135                 }
 136
 137                 if (copy) {
 138                         acl_entry_t new_entry;
 139
 140                         if (acl_create_entry(&copy, &new_entry) < 0)
 141                                 return -errno;
 142
 143                         if (acl_copy_entry(new_entry, i) < 0)
 144                                 return -errno;
 145
 146                         if (modify)
 147                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 148                                         return -errno;
 149                 }
 150
 151                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 152                 if (r < 0)
 153                         return -errno;
 154         }
 155
 156         *ret = TAKE_PTR(copy);
 157
 158         return !!*ret;
 159 }
 160
 161 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 162         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 163         bool changed = false;
 164         int r;
 165
 166         assert(fd >= 0);
 167         assert(st);
 168
 169         /* ACLs are not supported on symlinks, there's no point in trying */
 170         if (S_ISLNK(st->st_mode))
 171                 return 0;
 172
 173         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 174         if (r == -EOPNOTSUPP)
 175                 return 0;
 176         if (r < 0)
 177                 return r;
 178
 179         r = shift_acl(acl, shift, &shifted);
 180         if (r < 0)
 181                 return r;
 182         if (r > 0) {
 183                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 184                 if (r < 0)
 185                         return r;
 186
 187                 changed = true;
 188         }
 189
 190         if (S_ISDIR(st->st_mode)) {
 191                 acl_free(acl);
 192                 acl_free(shifted);
 193
 194                 acl = shifted = NULL;
 195
 196                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 197                 if (r < 0)
 198                         return r;
 199
 200                 r = shift_acl(acl, shift, &shifted);
 201                 if (r < 0)
 202                         return r;
 203                 if (r > 0) {
 204                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 205                         if (r < 0)
 206                                 return r;
 207
 208                         changed = true;
 209                 }
 210         }
 211
 212         return changed;
 213 }
 214
 215 #else
 216
 217 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 218         return 0;
 219 }
 220
 221 #endif
 222
 223 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 224         uid_t new_uid;
 225         gid_t new_gid;
 226         bool changed = false;
 227         int r;
 228
 229         assert(fd >= 0);
 230         assert(st);
 231
 232         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 233         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 234
 235         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 236                 return -EINVAL;
 237
 238         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 239                 if (name)
 240                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 241                 else
 242                         r = fchown(fd, new_uid, new_gid);
 243                 if (r < 0)
 244                         return -errno;
 245
 246                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 247                 if (name) {
 248                         if (!S_ISLNK(st->st_mode))
 249                                 r = fchmodat(fd, name, st->st_mode, 0);
 250                         else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
 251                                 r = 0;
 252                 } else
 253                         r = fchmod(fd, st->st_mode);
 254                 if (r < 0)
 255                         return -errno;
 256
 257                 changed = true;
 258         }
 259
 260         r = patch_acls(fd, name, st, shift);
 261         if (r < 0)
 262                 return r;
 263
 264         return r > 0 || changed;
 265 }
 266
 267 /*
 268  * Check if the filesystem is fully compatible with user namespaces or
 269  * UID/GID patching. Some filesystems in this list can be fully mounted inside
 270  * user namespaces, however their inodes may relate to host resources or only
 271  * valid in the global user namespace, therefore no patching should be applied.
 272  */
 273 static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
 274
 275         assert(sfs);
 276
 277         return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
 278                F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
 279                F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
 280                F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
 281                F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
 282                F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
 283                F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
 284                F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
 285                F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
 286                F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
 287                F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
 288                F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
 289                F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
 290                F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
 291                F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
 292                F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
 293 }
 294
 295 static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
 296         _cleanup_closedir_ DIR *d = NULL;
 297         bool changed = false;
 298         struct statfs sfs;
 299         int r;
 300
 301         assert(fd >= 0);
 302
 303         if (fstatfs(fd, &sfs) < 0)
 304                 return -errno;
 305
 306         /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
 307          * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
 308          * when we hit procfs, sysfs or some other special file systems. */
 309
 310         r = is_fs_fully_userns_compatible(&sfs);
 311         if (r < 0)
 312                 goto finish;
 313         if (r > 0) {
 314                 r = 0; /* don't recurse */
 315                 goto finish;
 316         }
 317
 318         /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
 319         if ((sfs.f_flags & ST_RDONLY) ||
 320             access_fd(fd, W_OK) == -EROFS)
 321                 goto read_only;
 322
 323         if (S_ISDIR(st->st_mode)) {
 324                 struct dirent *de;
 325
 326                 if (!donate_fd) {
 327                         int copy;
 328
 329                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 330                         if (copy < 0) {
 331                                 r = -errno;
 332                                 goto finish;
 333                         }
 334
 335                         fd = copy;
 336                         donate_fd = true;
 337                 }
 338
 339                 d = take_fdopendir(&fd);
 340                 if (!d) {
 341                         r = -errno;
 342                         goto finish;
 343                 }
 344
 345                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 346                         struct stat fst;
 347
 348                         if (dot_or_dot_dot(de->d_name))
 349                                 continue;
 350
 351                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 352                                 r = -errno;
 353                                 goto finish;
 354                         }
 355
 356                         if (S_ISDIR(fst.st_mode)) {
 357                                 int subdir_fd;
 358
 359                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 360                                 if (subdir_fd < 0) {
 361                                         r = -errno;
 362                                         goto finish;
 363
 364                                 }
 365
 366                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 367                                 if (r < 0)
 368                                         goto finish;
 369                                 if (r > 0)
 370                                         changed = true;
 371
 372                         } else {
 373                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 374                                 if (r < 0)
 375                                         goto finish;
 376                                 if (r > 0)
 377                                         changed = true;
 378                         }
 379                 }
 380         }
 381
 382         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 383          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 384          * tree is properly chown()ed already. */
 385         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 386         if (r == -EROFS)
 387                 goto read_only;
 388         if (r > 0)
 389                 changed = true;
 390
 391         r = changed;
 392         goto finish;
 393
 394 read_only:
 395         if (!is_toplevel) {
 396                 _cleanup_free_ char *name = NULL;
 397
 398                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 399                 (void) fd_get_path(fd, &name);
 400                 log_debug("Skipping read-only file or directory %s.", strna(name));
 401                 r = changed;
 402         }
 403
 404 finish:
 405         if (donate_fd)
 406                 safe_close(fd);
 407
 408         return r;
 409 }
 410
 411 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 412         struct stat st;
 413         int r;
 414
 415         assert(fd >= 0);
 416
 417         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 418          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 419          * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
 420          * UID within the container. */
 421
 422         if ((shift & 0xFFFF) != 0) {
 423                 /* We only support containers where the shift starts at a 2^16 boundary */
 424                 r = -EOPNOTSUPP;
 425                 goto finish;
 426         }
 427
 428         if (shift == UID_BUSY_BASE) {
 429                 r = -EINVAL;
 430                 goto finish;
 431         }
 432
 433         if (range != 0x10000) {
 434                 /* We only support containers with 16bit UID ranges for the patching logic */
 435                 r = -EOPNOTSUPP;
 436                 goto finish;
 437         }
 438
 439         if (fstat(fd, &st) < 0) {
 440                 r = -errno;
 441                 goto finish;
 442         }
 443
 444         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 445                 /* We only support containers where the uid/gid container ID match */
 446                 r = -EBADE;
 447                 goto finish;
 448         }
 449
 450         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 451          * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
 452         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 453                 return 0;
 454
 455         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 456          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 457          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 458
 459         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 460                 if (fchown(fd,
 461                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 462                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 463                         r = -errno;
 464                         goto finish;
 465                 }
 466         }
 467
 468         return recurse_fd(fd, donate_fd, &st, shift, true);
 469
 470 finish:
 471         if (donate_fd)
 472                 safe_close(fd);
 473
 474         return r;
 475 }
 476
 477 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 478         int fd;
 479
 480         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 481         if (fd < 0)
 482                 return -errno;
 483
 484         return fd_patch_uid_internal(fd, true, shift, range);
 485 }